CASTalk.com

We have a new toy: Athlon 64 X2 4400+ (Socket 939, 2.2GHz, 2*1MB L2
cache), Tyan S2865AG2NRF Tomcat K8E, on-board graphics (ATI Rage XL)
in text mode, 4GB PC3200 DDR SDRAM, 2 300GB hard disks spinning, 1
DVD-RW drive, Tagan TG480-U22 power supply.

Power consumption measured at the mains is (running Linux in text mode):

|------ power ------|
clock idle load 1 load 2
1000MHz 83W 93W 102W
1800MHz 86W 103W 121W
2000MHz 88W 109W 130W
2200MHz 92W 116W 143W

See
http://www.complang.tuwien.ac.at/anton/ ... ption.html
for results on other machines.

- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html

Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:

We have a new toy: Athlon 64 X2 4400+ (Socket 939, 2.2GHz, 2*1MB L2
cache), Tyan S2865AG2NRF Tomcat K8E, on-board graphics (ATI Rage XL)
in text mode, 4GB PC3200 DDR SDRAM, 2 300GB hard disks spinning, 1
DVD-RW drive, Tagan TG480-U22 power supply.

Power consumption measured at the mains is (running Linux in text mode):

|------ power ------|
clock idle load 1 load 2
1000MHz 83W 93W 102W
1800MHz 86W 103W 121W
2000MHz 88W 109W 130W
2200MHz 92W 116W 143W

See
http://www.complang.tuwien.ac.at/anton/ ... ption.html
for results on other machines.

Perhaps there is interest also in power consumption on a dual-dual G5:
see the second newsitem on <http://www.hardmac.com/news/2005-11-22/>.
Only the CPU power can be seen though, not total system power.

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

anton@mips.complang.tuwien.ac.at (Anton Ertl) writes:

We have a new toy: Athlon 64 X2 4400+ (Socket 939, 2.2GHz, 2*1MB L2
cache), Tyan S2865AG2NRF Tomcat K8E, on-board graphics (ATI Rage XL)
in text mode, 4GB PC3200 DDR SDRAM, 2 300GB hard disks spinning, 1
DVD-RW drive, Tagan TG480-U22 power supply.

Power consumption measured at the mains is (running Linux in text mode):
(added voltage according to powernow-k8 driver)

|------ power ------|
clock voltage idle load 1 load 2
1000MHz 1200mV 83W 93W 102W
1800MHz 1250mV 86W 103W 121W
2000MHz 1300mV 88W 109W 130W
2200MHz 1350mV 92W 116W 143W

Today we installed the other new toy:

Dual Opteron 270 (Socket 940, Dual-core 2GHz, 2*1MB L2 cache), Tyan
S2892 Thunder K8SE board with on-board graphics (ATI Rage XL), 8GB
PC3200 ECC RAM, 2 300GB SATA hard disks spinning, 1 DVD-RW drive,
Tagan TG-480-U22 power supply, running Linux-2.6.14.3.

clock voltage idle load 1 load 2 load 3 load 4
1000MHz 1100mV 121W 124W 127W 131W 134W
1800MHz 1350mV 161W 171W 181W 191W 202W
2000MHz 1350mV 167W 178W 190W 200W 212W
ondemand 121W 154W 165W/189W

Linux-2.6.14.3 seems to prefer to put the second process on the second
chip, so we usually got the 189W consumption with two non-nice
processes and the ondemand frequency governor; we got the 165W number
by starting three processes, then killing the middle one, so that both
processes run on the same chip, and the other chip idles.

This behaviour is not just bad for the power consumption, but also for
speed for many loads, because the ondemand governor needs some time
until it revs up the chip; OTOH, for longer-running, memory-intensive
loads, this behaviour helps speed.

The differences between the Athlon 64 X2 box and the Opteron box are
interesting: On the Opteron the frequency/voltage seems to have a much
higher influence on the idle power (23W/chip vs. 9W for the X2), but
the power consumption per loaded core is much smaller at the same
frequencies (3W vs. 10W @1000MHz, 11W vs. 21W @2000MHz). Looks like
AMD really does tune the different CPUs differently; alternatively,
these effects could come from different efficiencies of the various
voltage regulators at different loads, but the differences seem to be
a bit high for that (seriously inefficient voltage regulators would
probably need more cooling effort applied).

See
http://www.complang.tuwien.ac.at/anton/ ... ption.html
for results on other machines.

- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html

Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:

This behaviour is not just bad for the power consumption, but also for
speed for many loads, because the ondemand governor needs some time
until it revs up the chip; OTOH, for longer-running, memory-intensive
loads, this behaviour helps speed.

What latency do you see on frequency+voltage changes?

On my iMac G5 running MacOS X 10.4.3, in 8 tests I saw the change happen
in 0.1995168, 0.2588413, 0.2258136, 0.407833, 0.2753626, 0.2338377,
0.151981 and 0.146981 seconds from the time that the commandline testing
app had launched.

I have seen slides from IDF indicating Intel systems doing voltage
changes in milliseconds *before* they improved the speed. I wonder what
latency people see on shipping systems.

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

anton@mips.complang.tuwien.ac.at (Anton Ertl) writes:

1000MHz 1100mV 121W 124W 127W 131W 134W

FWIW a dual 250 (2.4Ghz running at 1Ghz) here requires about 136W
according to my really cheap wattmeter when idle.

1800MHz 1350mV 161W 171W 181W 191W 202W
2000MHz 1350mV 167W 178W 190W 200W 212W
ondemand 121W 154W 165W/189W

Linux-2.6.14.3 seems to prefer to put the second process on the second
chip, so we usually got the 189W consumption with two non-nice
processes and the ondemand frequency governor; we got the 165W number

ondemand is not a really good match for Opterons. It was designed
for CPUs with fast frequency transitions and powernow is comparatively
slow (there was a paper about it at OLS of a few years back if you
want details).

This behaviour is not just bad for the power consumption, but also for
speed for many loads, because the ondemand governor needs some time
until it revs up the chip; OTOH, for longer-running, memory-intensive
loads, this behaviour helps speed.

For loaded systems it's best though because it maximizes memory bandwidth.
Not necessarily for power saving.

-Andi

nospam@ab-katrinedal.dk (=?ISO-8859-1?Q?Niels_J=F8rgen_Kruse?=) writes:

I have seen slides from IDF indicating Intel systems doing voltage
changes in milliseconds *before* they improved the speed. I wonder what
latency people see on shipping systems.

The latency for Powernow is pretty bad, I think.

VIA C7 uses two PLL and can switch in one cycle, apparently.

Casper
--
Expressed in this posting are my opinions. They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.

nospam@ab-katrinedal.dk (=?ISO-8859-1?Q?Niels_J=F8rgen_Kruse?=) writes:

Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:

This behaviour is not just bad for the power consumption, but also for
speed for many loads, because the ondemand governor needs some time
until it revs up the chip; OTOH, for longer-running, memory-intensive
loads, this behaviour helps speed.

What latency do you see on frequency+voltage changes?

With the ondemand governor, I see the speedups happening at

0.70s
0.85s
0.95s
0.80s
0.74s
0.53s
0.95s
0.80s
0.62s
0.94s
0.46s
0.47s
0.62s
0.93s
0.77s

However, I have seen a Windows XP Cool'n'Quiet driver (on a system
with a Winchester Athlon 64) that seemed to react much faster (judging
from the frantic motion of the needle of the speedometer tool provided
with the driver; I did not do measurements), probably faster than
0.1s.

- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html

Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:

nospam@ab-katrinedal.dk (=?ISO-8859-1?Q?Niels_J=F8rgen_Kruse?=) writes:
Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:

This behaviour is not just bad for the power consumption, but also for
speed for many loads, because the ondemand governor needs some time
until it revs up the chip; OTOH, for longer-running, memory-intensive
loads, this behaviour helps speed.

What latency do you see on frequency+voltage changes?

With the ondemand governor, I see the speedups happening at

0.70s
0.85s
0.95s
0.80s
0.74s
0.53s
0.95s
0.80s
0.62s
0.94s
0.46s
0.47s
0.62s
0.93s
0.77s

This seems about in line with the latency I saw on my iMac when I first
recieved it, with the then current OS. I take it that the "ondemand
governor" is the component of Linux responsible for the speed
transitions? I would assume that this is the sort of thing that would
recieve tuning attention in Linux. Perhaps this feature is not so old?

However, I have seen a Windows XP Cool'n'Quiet driver (on a system
with a Winchester Athlon 64) that seemed to react much faster (judging
from the frantic motion of the needle of the speedometer tool provided
with the driver; I did not do measurements), probably faster than
0.1s.

On my first measurements, I would sometimes see an early switch to high
speed followed shortly by a drop back, I presume as a result of the
voltage droop caused by the higher power draw. So not all switches are
necessarily good.

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark

Anton Ertl wrote:

With the ondemand governor, I see the speedups happening at

Try changing /sys/devices/system/cpu/cpu0/cpufreq/ondemand/sampling_rate

However, the sampling_rate_min value is a read-only value, set by the
powernow-k8 code:

/* Take a crude guess here.
* That guess was in microseconds, so multiply with 1000 */
pol->cpuinfo.transition_latency =
(((data->rvo + 8) * data->vstable * VST_UNITS_20US)
+ (3 * (1 << data->irt) * 10)) * 1000;

The default value is twice the minimum value, so the best thing you can do
is to reduce the latency to half of that. The values come out of the ACPI
BIOS, so little control over that. The comment in the code says that this
is deprecated by AMD, so maybe there'll be a better way sometimes.

I've seen two different minimum rates on two different boards: 670000(us)
and 620000 (Gigabyte, Athlon 3200+, and ASUS Athlon x2 3800+).

--
Bernd Paysan
"If you want it done right, you have to do it yourself"
http://www.jwdt.com/~paysan/

Bernd Paysan <bernd.paysan@gmx.de> writes:

Anton Ertl wrote:
With the ondemand governor, I see the speedups happening at

Try changing /sys/devices/system/cpu/cpu0/cpufreq/ondemand/sampling_rate
....
I've seen two different minimum rates on two different boards: 670000(us)
and 620000 (Gigabyte, Athlon 3200+, and ASUS Athlon x2 3800+).

On the Opteron 270 box with the Tyan S2892SE the minimum sampling rate
is 280000. On my Asus K8V SE (with a Clawhammer Athlon 64 3200+),
it's 520000.

- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html